xref: /linux/arch/arm64/kvm/hyp/vhe/tlb.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2015 - ARM Ltd
4  * Author: Marc Zyngier <marc.zyngier@arm.com>
5  */
6 
7 #include <linux/irqflags.h>
8 
9 #include <asm/kvm_hyp.h>
10 #include <asm/kvm_mmu.h>
11 #include <asm/tlbflush.h>
12 
13 struct tlb_inv_context {
14 	struct kvm_s2_mmu	*mmu;
15 	unsigned long		flags;
16 	u64			tcr;
17 	u64			sctlr;
18 };
19 
20 static void enter_vmid_context(struct kvm_s2_mmu *mmu,
21 			       struct tlb_inv_context *cxt)
22 {
23 	struct kvm_vcpu *vcpu = kvm_get_running_vcpu();
24 	u64 val;
25 
26 	local_irq_save(cxt->flags);
27 
28 	if (vcpu && mmu != vcpu->arch.hw_mmu)
29 		cxt->mmu = vcpu->arch.hw_mmu;
30 	else
31 		cxt->mmu = NULL;
32 
33 	if (cpus_have_final_cap(ARM64_WORKAROUND_SPECULATIVE_AT)) {
34 		/*
35 		 * For CPUs that are affected by ARM errata 1165522 or 1530923,
36 		 * we cannot trust stage-1 to be in a correct state at that
37 		 * point. Since we do not want to force a full load of the
38 		 * vcpu state, we prevent the EL1 page-table walker to
39 		 * allocate new TLBs. This is done by setting the EPD bits
40 		 * in the TCR_EL1 register. We also need to prevent it to
41 		 * allocate IPA->PA walks, so we enable the S1 MMU...
42 		 */
43 		val = cxt->tcr = read_sysreg_el1(SYS_TCR);
44 		val |= TCR_EPD1_MASK | TCR_EPD0_MASK;
45 		write_sysreg_el1(val, SYS_TCR);
46 		val = cxt->sctlr = read_sysreg_el1(SYS_SCTLR);
47 		val |= SCTLR_ELx_M;
48 		write_sysreg_el1(val, SYS_SCTLR);
49 	}
50 
51 	/*
52 	 * With VHE enabled, we have HCR_EL2.{E2H,TGE} = {1,1}, and
53 	 * most TLB operations target EL2/EL0. In order to affect the
54 	 * guest TLBs (EL1/EL0), we need to change one of these two
55 	 * bits. Changing E2H is impossible (goodbye TTBR1_EL2), so
56 	 * let's flip TGE before executing the TLB operation.
57 	 *
58 	 * ARM erratum 1165522 requires some special handling (again),
59 	 * as we need to make sure both stages of translation are in
60 	 * place before clearing TGE. __load_stage2() already
61 	 * has an ISB in order to deal with this.
62 	 */
63 	__load_stage2(mmu, mmu->arch);
64 	val = read_sysreg(hcr_el2);
65 	val &= ~HCR_TGE;
66 	write_sysreg(val, hcr_el2);
67 	isb();
68 }
69 
70 static void exit_vmid_context(struct tlb_inv_context *cxt)
71 {
72 	/*
73 	 * We're done with the TLB operation, let's restore the host's
74 	 * view of HCR_EL2.
75 	 */
76 	write_sysreg(HCR_HOST_VHE_FLAGS, hcr_el2);
77 	isb();
78 
79 	/* ... and the stage-2 MMU context that we switched away from */
80 	if (cxt->mmu)
81 		__load_stage2(cxt->mmu, cxt->mmu->arch);
82 
83 	if (cpus_have_final_cap(ARM64_WORKAROUND_SPECULATIVE_AT)) {
84 		/* Restore the registers to what they were */
85 		write_sysreg_el1(cxt->tcr, SYS_TCR);
86 		write_sysreg_el1(cxt->sctlr, SYS_SCTLR);
87 	}
88 
89 	local_irq_restore(cxt->flags);
90 }
91 
92 void __kvm_tlb_flush_vmid_ipa(struct kvm_s2_mmu *mmu,
93 			      phys_addr_t ipa, int level)
94 {
95 	struct tlb_inv_context cxt;
96 
97 	dsb(ishst);
98 
99 	/* Switch to requested VMID */
100 	enter_vmid_context(mmu, &cxt);
101 
102 	/*
103 	 * We could do so much better if we had the VA as well.
104 	 * Instead, we invalidate Stage-2 for this IPA, and the
105 	 * whole of Stage-1. Weep...
106 	 */
107 	ipa >>= 12;
108 	__tlbi_level(ipas2e1is, ipa, level);
109 
110 	/*
111 	 * We have to ensure completion of the invalidation at Stage-2,
112 	 * since a table walk on another CPU could refill a TLB with a
113 	 * complete (S1 + S2) walk based on the old Stage-2 mapping if
114 	 * the Stage-1 invalidation happened first.
115 	 */
116 	dsb(ish);
117 	__tlbi(vmalle1is);
118 	dsb(ish);
119 	isb();
120 
121 	exit_vmid_context(&cxt);
122 }
123 
124 void __kvm_tlb_flush_vmid_ipa_nsh(struct kvm_s2_mmu *mmu,
125 				  phys_addr_t ipa, int level)
126 {
127 	struct tlb_inv_context cxt;
128 
129 	dsb(nshst);
130 
131 	/* Switch to requested VMID */
132 	enter_vmid_context(mmu, &cxt);
133 
134 	/*
135 	 * We could do so much better if we had the VA as well.
136 	 * Instead, we invalidate Stage-2 for this IPA, and the
137 	 * whole of Stage-1. Weep...
138 	 */
139 	ipa >>= 12;
140 	__tlbi_level(ipas2e1, ipa, level);
141 
142 	/*
143 	 * We have to ensure completion of the invalidation at Stage-2,
144 	 * since a table walk on another CPU could refill a TLB with a
145 	 * complete (S1 + S2) walk based on the old Stage-2 mapping if
146 	 * the Stage-1 invalidation happened first.
147 	 */
148 	dsb(nsh);
149 	__tlbi(vmalle1);
150 	dsb(nsh);
151 	isb();
152 
153 	exit_vmid_context(&cxt);
154 }
155 
156 void __kvm_tlb_flush_vmid_range(struct kvm_s2_mmu *mmu,
157 				phys_addr_t start, unsigned long pages)
158 {
159 	struct tlb_inv_context cxt;
160 	unsigned long stride;
161 
162 	/*
163 	 * Since the range of addresses may not be mapped at
164 	 * the same level, assume the worst case as PAGE_SIZE
165 	 */
166 	stride = PAGE_SIZE;
167 	start = round_down(start, stride);
168 
169 	dsb(ishst);
170 
171 	/* Switch to requested VMID */
172 	enter_vmid_context(mmu, &cxt);
173 
174 	__flush_s2_tlb_range_op(ipas2e1is, start, pages, stride,
175 				TLBI_TTL_UNKNOWN);
176 
177 	dsb(ish);
178 	__tlbi(vmalle1is);
179 	dsb(ish);
180 	isb();
181 
182 	exit_vmid_context(&cxt);
183 }
184 
185 void __kvm_tlb_flush_vmid(struct kvm_s2_mmu *mmu)
186 {
187 	struct tlb_inv_context cxt;
188 
189 	dsb(ishst);
190 
191 	/* Switch to requested VMID */
192 	enter_vmid_context(mmu, &cxt);
193 
194 	__tlbi(vmalls12e1is);
195 	dsb(ish);
196 	isb();
197 
198 	exit_vmid_context(&cxt);
199 }
200 
201 void __kvm_flush_cpu_context(struct kvm_s2_mmu *mmu)
202 {
203 	struct tlb_inv_context cxt;
204 
205 	/* Switch to requested VMID */
206 	enter_vmid_context(mmu, &cxt);
207 
208 	__tlbi(vmalle1);
209 	asm volatile("ic iallu");
210 	dsb(nsh);
211 	isb();
212 
213 	exit_vmid_context(&cxt);
214 }
215 
216 void __kvm_flush_vm_context(void)
217 {
218 	dsb(ishst);
219 	__tlbi(alle1is);
220 	dsb(ish);
221 }
222 
223 /*
224  * TLB invalidation emulation for NV. For any given instruction, we
225  * perform the following transformtions:
226  *
227  * - a TLBI targeting EL2 S1 is remapped to EL1 S1
228  * - a non-shareable TLBI is upgraded to being inner-shareable
229  * - an outer-shareable TLBI is also mapped to inner-shareable
230  * - an nXS TLBI is upgraded to XS
231  */
232 int __kvm_tlbi_s1e2(struct kvm_s2_mmu *mmu, u64 va, u64 sys_encoding)
233 {
234 	struct tlb_inv_context cxt;
235 	int ret = 0;
236 
237 	/*
238 	 * The guest will have provided its own DSB ISHST before trapping.
239 	 * If it hasn't, that's its own problem, and we won't paper over it
240 	 * (plus, there is plenty of extra synchronisation before we even
241 	 * get here...).
242 	 */
243 
244 	if (mmu)
245 		enter_vmid_context(mmu, &cxt);
246 
247 	switch (sys_encoding) {
248 	case OP_TLBI_ALLE2:
249 	case OP_TLBI_ALLE2IS:
250 	case OP_TLBI_ALLE2OS:
251 	case OP_TLBI_VMALLE1:
252 	case OP_TLBI_VMALLE1IS:
253 	case OP_TLBI_VMALLE1OS:
254 	case OP_TLBI_ALLE2NXS:
255 	case OP_TLBI_ALLE2ISNXS:
256 	case OP_TLBI_ALLE2OSNXS:
257 	case OP_TLBI_VMALLE1NXS:
258 	case OP_TLBI_VMALLE1ISNXS:
259 	case OP_TLBI_VMALLE1OSNXS:
260 		__tlbi(vmalle1is);
261 		break;
262 	case OP_TLBI_VAE2:
263 	case OP_TLBI_VAE2IS:
264 	case OP_TLBI_VAE2OS:
265 	case OP_TLBI_VAE1:
266 	case OP_TLBI_VAE1IS:
267 	case OP_TLBI_VAE1OS:
268 	case OP_TLBI_VAE2NXS:
269 	case OP_TLBI_VAE2ISNXS:
270 	case OP_TLBI_VAE2OSNXS:
271 	case OP_TLBI_VAE1NXS:
272 	case OP_TLBI_VAE1ISNXS:
273 	case OP_TLBI_VAE1OSNXS:
274 		__tlbi(vae1is, va);
275 		break;
276 	case OP_TLBI_VALE2:
277 	case OP_TLBI_VALE2IS:
278 	case OP_TLBI_VALE2OS:
279 	case OP_TLBI_VALE1:
280 	case OP_TLBI_VALE1IS:
281 	case OP_TLBI_VALE1OS:
282 	case OP_TLBI_VALE2NXS:
283 	case OP_TLBI_VALE2ISNXS:
284 	case OP_TLBI_VALE2OSNXS:
285 	case OP_TLBI_VALE1NXS:
286 	case OP_TLBI_VALE1ISNXS:
287 	case OP_TLBI_VALE1OSNXS:
288 		__tlbi(vale1is, va);
289 		break;
290 	case OP_TLBI_ASIDE1:
291 	case OP_TLBI_ASIDE1IS:
292 	case OP_TLBI_ASIDE1OS:
293 	case OP_TLBI_ASIDE1NXS:
294 	case OP_TLBI_ASIDE1ISNXS:
295 	case OP_TLBI_ASIDE1OSNXS:
296 		__tlbi(aside1is, va);
297 		break;
298 	case OP_TLBI_VAAE1:
299 	case OP_TLBI_VAAE1IS:
300 	case OP_TLBI_VAAE1OS:
301 	case OP_TLBI_VAAE1NXS:
302 	case OP_TLBI_VAAE1ISNXS:
303 	case OP_TLBI_VAAE1OSNXS:
304 		__tlbi(vaae1is, va);
305 		break;
306 	case OP_TLBI_VAALE1:
307 	case OP_TLBI_VAALE1IS:
308 	case OP_TLBI_VAALE1OS:
309 	case OP_TLBI_VAALE1NXS:
310 	case OP_TLBI_VAALE1ISNXS:
311 	case OP_TLBI_VAALE1OSNXS:
312 		__tlbi(vaale1is, va);
313 		break;
314 	case OP_TLBI_RVAE2:
315 	case OP_TLBI_RVAE2IS:
316 	case OP_TLBI_RVAE2OS:
317 	case OP_TLBI_RVAE1:
318 	case OP_TLBI_RVAE1IS:
319 	case OP_TLBI_RVAE1OS:
320 	case OP_TLBI_RVAE2NXS:
321 	case OP_TLBI_RVAE2ISNXS:
322 	case OP_TLBI_RVAE2OSNXS:
323 	case OP_TLBI_RVAE1NXS:
324 	case OP_TLBI_RVAE1ISNXS:
325 	case OP_TLBI_RVAE1OSNXS:
326 		__tlbi(rvae1is, va);
327 		break;
328 	case OP_TLBI_RVALE2:
329 	case OP_TLBI_RVALE2IS:
330 	case OP_TLBI_RVALE2OS:
331 	case OP_TLBI_RVALE1:
332 	case OP_TLBI_RVALE1IS:
333 	case OP_TLBI_RVALE1OS:
334 	case OP_TLBI_RVALE2NXS:
335 	case OP_TLBI_RVALE2ISNXS:
336 	case OP_TLBI_RVALE2OSNXS:
337 	case OP_TLBI_RVALE1NXS:
338 	case OP_TLBI_RVALE1ISNXS:
339 	case OP_TLBI_RVALE1OSNXS:
340 		__tlbi(rvale1is, va);
341 		break;
342 	case OP_TLBI_RVAAE1:
343 	case OP_TLBI_RVAAE1IS:
344 	case OP_TLBI_RVAAE1OS:
345 	case OP_TLBI_RVAAE1NXS:
346 	case OP_TLBI_RVAAE1ISNXS:
347 	case OP_TLBI_RVAAE1OSNXS:
348 		__tlbi(rvaae1is, va);
349 		break;
350 	case OP_TLBI_RVAALE1:
351 	case OP_TLBI_RVAALE1IS:
352 	case OP_TLBI_RVAALE1OS:
353 	case OP_TLBI_RVAALE1NXS:
354 	case OP_TLBI_RVAALE1ISNXS:
355 	case OP_TLBI_RVAALE1OSNXS:
356 		__tlbi(rvaale1is, va);
357 		break;
358 	default:
359 		ret = -EINVAL;
360 	}
361 	dsb(ish);
362 	isb();
363 
364 	if (mmu)
365 		exit_vmid_context(&cxt);
366 
367 	return ret;
368 }
369